Autonomous navigation for guide following in crowded indoor environments

Abstract

The requirements for assisted living are rapidly changing as the number of elderly patients over the age of 60 continues to increase. This rise places a high level of stress on nurse practitioners who must care for more patients than they are capable. As this trend is expected to continue, new technology will be required to help care for patients. Mobile robots present an opportunity to help alleviate the stress on nurse practitioners by monitoring and performing remedial tasks for elderly patients. In order to produce mobile robots with the ability to perform these tasks, however, many challenges must be overcome. The hospital environment requires a high level of safety to prevent patient injury. Any facility that uses mobile robots, therefore, must be able to ensure that no harm will come to patients whilst in a care environment. This requires the robot to build a high level of understanding about the environment and the people with close proximity to the robot. Hitherto, most mobile robots have used vision-based sensors or 2D laser range finders. 3D time-of-flight sensors have recently been introduced and provide dense 3D point clouds of the environment at real-time frame rates. This provides mobile robots with previously unavailable dense information in real-time. I investigate the use of time-of-flight cameras for mobile robot navigation in crowded environments in this thesis. A unified framework to allow the robot to follow a guide through an indoor environment safely and efficiently is presented. Each component of the framework is analyzed in detail, with real-world scenarios illustrating its practical use. Time-of-flight cameras are relatively new sensors and, therefore, have inherent problems that must be overcome to receive consistent and accurate data. I propose a novel and practical probabilistic framework to overcome many of the inherent problems in this thesis. The framework fuses multiple depth maps with color information forming a reliable and consistent view of the world. In order for the robot to interact with the environment, contextual information is required. To this end, I propose a region-growing segmentation algorithm to group points based on surface characteristics, surface normal and surface curvature. The segmentation process creates a distinct set of surfaces, however, only a limited amount of contextual information is available to allow for interaction. Therefore, a novel classifier is proposed using spherical harmonics to differentiate people from all other objects. The added ability to identify people allows the robot to find potential candidates to follow. However, for safe navigation, the robot must continuously track all visible objects to obtain positional and velocity information. A multi-object tracking system is investigated to track visible objects reliably using multiple cues, shape and color. The tracking system allows the robot to react to the dynamic nature of people by building an estimate of the motion flow. This flow provides the robot with the necessary information to determine where and at what speeds it is safe to drive. In addition, a novel search strategy is proposed to allow the robot to recover a guide who has left the field-of-view. To achieve this, a search map is constructed with areas of the environment ranked according to how likely they are to reveal the guide’s true location. Then, the robot can approach the most likely search area to recover the guide. Finally, all components presented are joined to follow a guide through an indoor environment. The results achieved demonstrate the efficacy of the proposed components.